Source corpus
Consecutive 10-Ks form the comparison layer. Earlier-disclosure checks use the frozen 8-K and first-substantive-headline corpora.
Methodology and trust
Every conclusion should be inspectable from the original evidence.
Plain language comes first. Technical detail is available when a reviewer needs to audit source identity, timing, extraction, or classification.
How the demonstration works
Nine controls protect the evidence chain
The frozen data and classifications were not altered to improve the visual story.
Consecutive comparison
Matched sections from one annual filing are compared with the issuer’s next annual filing.
Information timing
Only information available strictly before the current filing can affect novelty or classification.
Same-day exclusion
Same-day news is excluded to avoid uncertain ordering around the filing timestamp.
Evidence offsets
Every retained change has prior and current excerpt locations. All 1,990 offsets were validated.
Confidence
A fixed deterministic score records strength of match and classification; it is not a probability of return.
Novelty categories
Previously disclosed, genuinely new relative to the study sources, partially anticipated, or unclear.
Deterministic sampling
The 150 issuer-disjoint filing pairs and thresholds were frozen before the final bounded tests.
Outcome-independent cases
Eight cases were chosen for evidence diversity and clarity—not later market performance.
Trust checklist
What a reviewer can verify
Trust is a product feature only when the evidence is visible and the boundary is explicit.
- ✓ Original source visible
- ✓ Filing timestamp preserved
- ✓ Prior and current excerpts shown
- ✓ Evidence offsets validated
- ✓ Earlier disclosures checked
- ✓ Future outcomes excluded
- ✓ Analyst remains in control
- ✓ Zero point-in-time violations
Novelty, in plain language
“Genuinely new” is a bounded search result—not a claim about the entire world.
The label means the frozen search found no earlier match in the covered 8-K and pre-filing headline sources.
Previously disclosed
A relevant earlier 8-K or strictly pre-filing headline exists.
Genuinely new
No earlier match exists within the frozen filings and headlines available in this study.
Partially anticipated
Earlier sources contain part, but not all, of the current disclosure.
Unclear
The system cannot resolve novelty confidently; analyst review is required.
Interpretation layers
Evidence, system output, judgment, and later outcomes stay separate.
This separation prevents future information from leaking into a filing-time claim.
Available at filing time
Source evidence
Prior/current filings, strictly earlier 8-Ks, and pre-filing headlines.
System interpretation
Structured triage
Change, category, direction, materiality, novelty, confidence, and evidence links.
Analyst interpretation
Final judgment
Relevance, economic significance, model updates, and the conclusion shared with the team.
Observed later
Ex-post outcome
Subsequent market behavior, labeled separately and never used for classification or case selection.
Experimental Research Idea Engine
How Research Ideas Are Grounded
The eight-case system is suitable for demonstration, but broader generation remains blocked until formal blinded human review confirms grounding quality and usefulness.
- Full prior-filing searchSearch complete prior sections and filing context.
- Earlier 8-K and news checksUse only point-in-time eligible sources.
- Atomic claim decompositionTest each factual proposition separately.
- Financial-quantity normalizationAssign amounts their economic roles and bases.
- Novelty verificationFind the earliest occurrence and state what changed.
- Counterevidence retrievalActively seek evidence that weakens the narrative.
- Hypothesis generationRequire testable predictions and falsification conditions.
- Skeptical reviewReject incomplete context and unsupported mechanisms.
- Analyst decisionAccept, edit, reject, or escalate with human accountability.
- ✓ Complete prior filing searched
- ✓ Earlier 8-Ks checked
- ✓ Prior news checked
- ✓ Exact evidence shown
- ✓ Facts separated from inference
- ✓ Financial quantities assigned correct roles
- ✓ Counterevidence shown
- ✓ Falsification conditions required
- ✓ Unsupported mechanisms rejected
- ✓ No position sizing
- ✓ Analyst review required
Limitations
What this methodology cannot establish
The product demonstration is operational; the risk analysis remains inconclusive.
Not established
- Alpha or trading value
- Causality
- Stable disclosure-risk relationship
- Customer adoption or willingness to pay
- Institutional validation
Coverage limits
- Novelty is corpus-bounded
- Headline bodies are excluded
- One SEC 8-K parse remains unresolved
- Classification is deterministic triage
- Sector is mapped only for curated cases
Defensible today
- 150 issuer-disjoint pairs
- 995 evidence-backed changes
- 629 high-confidence records
- 1,990 verified offsets
- Zero timing violations
Technical corpus detail
The secondary archive contains 296,119 validated parsed SEC 8-K records from 2012–2024 and 731,682 frozen first-substantive-story headlines from 2015–2024. Article bodies are not redistributed. Annual partitions load only on request.